def test_c2st_and_map_snl_on_linearGaussian_different(num_dim: int,
prior_str: str):
"""Test SNL on linear Gaussian, comparing to ground truth posterior via c2st.
Args:
num_dim: parameter dimension of the gaussian model
prior_str: one of "gaussian" or "uniform"
"""
num_samples = 500
num_simulations = 3000
trials_to_test = [1]
# likelihood_mean will be likelihood_shift+theta
likelihood_shift = -1.0 * ones(num_dim)
# Use increased cov to avoid too small posterior cov for many trials.
likelihood_cov = 0.8 * eye(num_dim)
if prior_str == "gaussian":
prior_mean = zeros(num_dim)
prior_cov = eye(num_dim)
prior = MultivariateNormal(loc=prior_mean, covariance_matrix=prior_cov)
else:
prior = utils.BoxUniform(-2.0 * ones(num_dim), 2.0 * ones(num_dim))
simulator, prior = prepare_for_sbi(
lambda theta: linear_gaussian(theta, likelihood_shift, likelihood_cov),
prior)
density_estimator = likelihood_nn("maf", num_transforms=3)
inference = SNLE(density_estimator=density_estimator,
show_progress_bars=False)
theta, x = simulate_for_sbi(simulator,
prior,
num_simulations,
simulation_batch_size=10000)
likelihood_estimator = inference.append_simulations(theta, x).train()
# Test inference amortized over trials.
for num_trials in trials_to_test:
x_o = zeros((num_trials, num_dim))
if prior_str == "gaussian":
gt_posterior = true_posterior_linear_gaussian_mvn_prior(
x_o, likelihood_shift, likelihood_cov, prior_mean, prior_cov)
target_samples = gt_posterior.sample((num_samples, ))
elif prior_str == "uniform":
target_samples = samples_true_posterior_linear_gaussian_uniform_prior(
x_o,
likelihood_shift,
likelihood_cov,
prior=prior,
num_samples=num_samples,
)
else:
raise ValueError(f"Wrong prior_str: '{prior_str}'.")
potential_fn, theta_transform = likelihood_estimator_based_potential(
prior=prior, likelihood_estimator=likelihood_estimator, x_o=x_o)
posterior = MCMCPosterior(
proposal=prior,
potential_fn=potential_fn,
theta_transform=theta_transform,
method="slice_np_vectorized",
thin=5,
num_chains=5,
)
samples = posterior.sample(sample_shape=(num_samples, ))
# Check performance based on c2st accuracy.
check_c2st(samples,
target_samples,
alg=f"snle_a-{prior_str}-prior-{num_trials}-trials")
map_ = posterior.map(num_init_samples=1_000,
init_method="proposal",
show_progress_bars=False)
# TODO: we do not have a test for SNL log_prob(). This is because the output
# TODO: density is not normalized, so KLd does not make sense.
if prior_str == "uniform":
# Check whether the returned probability outside of the support is zero.
posterior_prob = get_prob_outside_uniform_prior(
posterior, prior, num_dim)
assert (
posterior_prob == 0.0
), "The posterior probability outside of the prior support is not zero"
assert ((map_ - ones(num_dim))**2).sum() < 0.5
else:
assert ((map_ - gt_posterior.mean)**2).sum() < 0.5
def test_c2st_sre_variants_on_linearGaussian(
num_dim: int,
num_trials: int,
prior_str: str,
method_str: str,
):
"""Test c2st accuracy of inference with SRE on linear Gaussian model.
Args:
num_dim: parameter dimension of the gaussian model
prior_str: one of "gaussian" or "uniform"
"""
x_o = zeros(num_trials, num_dim)
num_samples = 500
num_simulations = 2600 if num_trials == 1 else 40500
# `likelihood_mean` will be `likelihood_shift + theta`.
likelihood_shift = -1.0 * ones(num_dim)
likelihood_cov = 0.8 * eye(num_dim)
if prior_str == "gaussian":
prior_mean = zeros(num_dim)
prior_cov = eye(num_dim)
prior = MultivariateNormal(loc=prior_mean, covariance_matrix=prior_cov)
else:
prior = utils.BoxUniform(-2.0 * ones(num_dim), 2.0 * ones(num_dim))
def simulator(theta):
return linear_gaussian(theta, likelihood_shift, likelihood_cov)
simulator, prior = prepare_for_sbi(simulator, prior)
kwargs = dict(
classifier="resnet",
show_progress_bars=False,
)
inference = SNRE_B(**kwargs) if method_str == "sre" else AALR(**kwargs)
# Should use default `num_atoms=10` for SRE; `num_atoms=2` for AALR
theta, x = simulate_for_sbi(
simulator, prior, num_simulations, simulation_batch_size=50
)
ratio_estimator = inference.append_simulations(theta, x).train()
potential_fn, theta_transform = ratio_estimator_based_potential(
ratio_estimator=ratio_estimator, prior=prior, x_o=x_o
)
posterior = MCMCPosterior(
potential_fn=potential_fn,
theta_transform=theta_transform,
proposal=prior,
method="slice_np_vectorized",
thin=5,
num_chains=5,
)
samples = posterior.sample(sample_shape=(num_samples,))
# Get posterior samples.
if prior_str == "gaussian":
gt_posterior = true_posterior_linear_gaussian_mvn_prior(
x_o, likelihood_shift, likelihood_cov, prior_mean, prior_cov
)
target_samples = gt_posterior.sample((num_samples,))
else:
target_samples = samples_true_posterior_linear_gaussian_uniform_prior(
x_o, likelihood_shift, likelihood_cov, prior=prior, num_samples=num_samples
)
# Check performance based on c2st accuracy.
check_c2st(
samples, target_samples, alg=f"sre-{prior_str}-{method_str}-{num_trials}trials"
)
map_ = posterior.map(num_init_samples=1_000, init_method="proposal")
# Checks for log_prob()
if prior_str == "gaussian" and method_str == "aalr":
# For the Gaussian prior, we compute the KLd between ground truth and
# posterior. We can do this only if the classifier_loss was as described in
# Hermans et al. 2020 ('aalr') since Durkan et al. 2020 version only allows
# evaluation up to a constant.
# For the Gaussian prior, we compute the KLd between ground truth and posterior
dkl = get_dkl_gaussian_prior(
posterior, x_o, likelihood_shift, likelihood_cov, prior_mean, prior_cov
)
max_dkl = 0.15
assert (
dkl < max_dkl
), f"KLd={dkl} is more than 2 stds above the average performance."
assert ((map_ - gt_posterior.mean) ** 2).sum() < 0.5
if prior_str == "uniform":
# Check whether the returned probability outside of the support is zero.
posterior_prob = get_prob_outside_uniform_prior(posterior, prior, num_dim)
assert (
posterior_prob == 0.0
), "The posterior probability outside of the prior support is not zero"
assert ((map_ - ones(num_dim)) ** 2).sum() < 0.5